Gateway documentation

LLMGuard exposes an OpenAI-compatible API. Any OpenAI SDK works by changing the base URL and the API key.

Quickstart

Connect a provider in Settings → Providers (your own key, encrypted at rest).
Create an application and an LLMGuard API key (shown once).
Send requests to POST /v1/chat/completions.

curl $GATEWAY_URL/v1/chat/completions \
  -H "Authorization: Bearer llmg_live_…" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-4o-mini","messages":[{"role":"user","content":"Hello"}]}'

Models

Models are namespaced as provider/model: openai/gpt-4.1, anthropic/claude-sonnet-5-5, google/gemini-2.5-flash, mistral/mistral-large-latest, deepseek/deepseek-chat. GET /v1/models lists the models available to your key.

Streaming

Set "stream": true to receive Server-Sent Events in the OpenAI chunk format, ending with data: [DONE]. Add stream_options.include_usage to receive a final usage chunk. Usage is recorded even if the client disconnects mid-stream.

Response headers

x-request-id — correlation id, searchable in Monitoring → Requests
x-uaig-dlp-action — ALLOW, LOG or MASK
x-uaig-provider — provider that served the request
x-uaig-cost — estimated cost (non-streaming)
x-uaig-quota-usage-percent — highest consumption among applicable limits

Errors

Errors use an OpenAI-compatible envelope and never contain provider credentials or stack traces.

{ "error": { "code": "DLP_BLOCKED", "type": "permission_error",
             "message": "Request blocked by data loss prevention: …",
             "details": { "categories": ["FINANCIAL"], "rules": ["Credit card number"] },
             "request_id": "req_…" } }

400	INVALID_REQUEST	Malformed body, unknown model namespace, provider-side validation error
401	INVALID_API_KEY	Missing, malformed, unknown, revoked or expired LLMGuard API key
403	POLICY_BLOCKED	A policy forbids this provider/model/size (details.violation)
403	DLP_BLOCKED	Sensitive data detected (details.categories, details.rules — never the values)
429	RATE_LIMITED	Upstream provider rate limit, or too many invalid key attempts
429	QUOTA_EXCEEDED	A quota or budget with BLOCK enforcement is exhausted (Retry-After header)
502	PROVIDER_ERROR	Provider rejected the organization credentials or returned an unexpected error
503	PROVIDER_UNAVAILABLE	Provider unreachable or timed out
503	PROVIDER_NOT_CONFIGURED	The provider is not connected for your organization

DLP behaviour

All text fields (message content, content parts, tool-call arguments) are normalized (Unicode NFKC, zero-width characters removed) and scanned before forwarding. BLOCK rejects the request; MASK replaces matches with tokens such as [REDACTED_EMAIL]; LOG records a detection and forwards unchanged. Detection logs never contain matched values.

Quotas & budgets

Quotas (requests/minute, requests/day, tokens/day, tokens/month, cost/day, cost/month) and monthly budgets apply to the organization, a team, an application or a key. Request counts are enforced atomically; token and cost limits are checked before each request and charged with the provider-reported usage afterwards. Periods are computed in UTC.

Webhooks

Alerts (budget/quota thresholds, high-severity DLP, provider incidents) are delivered as JSON POSTs signed with x-uaig-signature: t=<unix>,v1=<hex HMAC-SHA256(secret, "t.body")>.